Delta-Learning approach combined with the cluster Gutzwiller approximation for strongly correlated bosonic systems

The cluster Gutzwiller method is widely used to study the strongly correlated bosonic systems, owing to its ability to provide a more precise description of quantum fluctuations. However, its utility is limited by the exponential increase in computational complexity as the cluster size grows. To overcome this limitation, we propose an artificial intelligence-based method known as $Δ$-Learning. This approach constructs a predictive model by learning the discrepancies between lower-precision (small cluster sizes) and high-precision (large cluster sizes) implementations of the cluster Gutzwiller method, requiring only a small number of training samples. Using this predictive model, we can effectively forecast the outcomes of high-precision methods with high accuracy. Applied to various Bose-Hubbard models, the $Δ$-Learning method effectively predicts phase diagrams while significantly reducing the computational resources and time. Furthermore, we have compared the predictive accuracy of $Δ$-Learning with other direct learning methods and found that $Δ$-Learning exhibits superior performance in scenarios with limited training data. Therefore, when combined with the cluster Gutzwiller approximation, the $Δ$-Learning approach offers a computationally efficient and accurate method for studying phase transitions in large, complex bosonic systems.

💡 Research Summary

**
The paper addresses a long‑standing bottleneck in the numerical study of strongly correlated bosonic systems: the exponential growth of computational cost when the cluster size is increased in the cluster Gutzwiller (CG) method. While CG improves upon single‑site mean‑field theory by treating a super‑cell of lattice sites exactly, the Hilbert space dimension scales as the product of local Fock spaces, quickly exhausting memory and CPU resources for clusters larger than a few sites. To overcome this limitation, the authors import the Δ‑Learning (Delta‑Learning) paradigm from quantum chemistry. Δ‑Learning builds a high‑accuracy predictor by learning the difference Δ(y)=y_high−y_low between a low‑precision baseline (small‑cluster CG) and a high‑precision target (large‑cluster CG). Only a handful of training points—typically four—are required because the correction Δ(y) is a smooth, low‑amplitude function that is much easier to approximate than the full observable.

Two machine‑learning models are examined: support‑vector machines (SVM) and back‑propagation neural networks (BPNN). Performance is measured by the mean absolute percentage error (MAPE) as a function of the number of training samples. The results show that Δ‑Learning consistently outperforms direct learning (which attempts to map the raw parameters directly to the high‑precision output) for small training sets. In particular, SVM‑based Δ‑Learning achieves sub‑1 % MAPE with as few as three training points and stabilises below 0.5 % error for four or more points, whereas BPNN requires more data to reach comparable accuracy. The superiority of SVM is attributed to its kernel‑based capacity to capture non‑linear relationships in high‑dimensional feature space while avoiding over‑fitting on limited data.

The methodology is applied to three representative Bose‑Hubbard models:

1. Square lattice Bose‑Hubbard model – baseline calculations are performed with a 2×2 cluster, while the target high‑precision data come from 3×3 and 4×4 clusters. The SVM‑Δ‑Learning predictions (plotted as green and red circles) lie almost exactly on the phase boundaries obtained by direct CG (solid lines).

2. Hexagonal (non‑Bravais) lattice – similar training on 2×2 baselines and 3×6 / 4×3 target clusters yields predictions that match the CG results for both the superfluid–Mott insulator transition and the more subtle density‑wave phases.

3. Bipartite superlattice – a more complex unit cell with alternating on‑site interactions is studied. Again, with only four training points, the Δ‑Learning model reproduces the intricate phase diagram, including both first‑order and second‑order transitions, as verified against large‑cluster CG calculations.

Across all cases, the computational savings are dramatic. The baseline CG calculations on small clusters are inexpensive, and the ML inference step is essentially instantaneous. Compared with performing a full CG sweep on large clusters, the combined Δ‑Learning approach reduces total CPU time by roughly an order of magnitude while preserving the same level of quantitative accuracy.

The authors also discuss the trade‑off inherent to Δ‑Learning: the need to compute a baseline adds a modest overhead, but this cost is negligible relative to the savings from avoiding large‑cluster diagonalizations. They argue that the physical plausibility of the baseline (it already respects the underlying Hamiltonian and symmetries) helps constrain the ML model, leading to better generalisation than direct learning, which must infer the full mapping from sparse data.

In the concluding section, several future directions are proposed. One is to replace the CG baseline with other efficient approximations (e.g., variational cluster approaches, Gutzwiller‑DMRG hybrids) to further improve scalability. Another is to extend the Δ‑Learning framework to multi‑parameter phase diagrams (temperature, disorder, external fields) using multi‑output regression or active‑learning strategies to select the most informative training points. Transfer learning across different lattice geometries is also suggested as a way to leverage previously trained models for new systems.

Overall, the paper demonstrates that integrating Δ‑Learning with the cluster Gutzwiller method provides a powerful, data‑efficient route to high‑precision phase‑diagram calculations for strongly correlated bosonic lattices. It bridges the gap between physically grounded many‑body approximations and modern machine‑learning techniques, delivering both computational efficiency and robust predictive performance.